A flagship AI model usually means a flagship price. The companies that can't pay step aside. Xiaomi, a phone maker, just released a model that claims to match the frontrunners for a fraction of the cost. And it gave the weights away.
The release: MiMo-V2.6
On September 22, 2026, Xiaomi (小米) released and open-sourced the MiMo-V2.6 series: a flagship MiMo-V2.6-Pro and an efficient MiMo-V2.6-Flash. Both natively accept text, image, video, and audio — what Xiaomi calls full-modal (全模态) input.
Reported specs for the Pro:
- Over 1 trillion total parameters, with about 42B active (MoE-style sparse activation)
- 1 million-token context
- Positioned as the only open-source full-modal Pro model in its tier
Xiaomi says the family is released under a permissive license (the project site lists MIT for the MiMo line), and that it also open-sourced the reinforcement-learning know-how behind the models.
The claim that matters: cost per task
The headline is not just capability but economics. On Artificial Analysis's composite intelligence index, Xiaomi says MiMo-V2.6-Pro scored 46 points, ranking first among open-weight models and first among domestic Chinese models — ahead of Kimi K3 and Qwen3.8 Max.
More striking is the cost column. Xiaomi cites a measured unit task cost of US$0.13 (about HK$1.0), which it claims is one-twentieth to one-sixtieth the cost of overseas models at similar intelligence. If even roughly true, it redraws the "smart-but-affordable" line for self-hosted and API users alike.
What it's actually for
MiMo-V2.6 is pitched at real work, not demos:
- Complex reasoning and long-horizon tasks
- Multi-agent collaboration and large coding projects
- Office, web, and frontend design; video editing; 3D and music generation
- Computer control, cybersecurity, and research
The most interesting pilot is scientific. Xiaomi says MiMo-V2.6-Pro acted as a "Co-Scientist," helping its materials team design and screen metal-organic framework (MOF) candidates for adsorbing PFAS — the "forever chemicals" — in a real internal research loop.
The bigger Xiaomi bet
MiMo is not a side project. It powers Xiaomi's "human-car-home" (人车家) ecosystem: the XiaoAI assistant, HyperOS, and smart cockpits. Baidu Baike notes Xiaomi plans to invest at least ¥60 billion in AI over three years — roughly US$8.5B / HK$66B — and that its on-device MiMo model passed national large-model filing in July 2026.
That context explains the strategy: a phone-and-device company wants models that run on its own hardware and don't bleed margin to outside APIs.
How it compares
- vs. closed frontiers: Xiaomi claims parity on cost-adjusted intelligence, not raw score.
- vs. other open weights: it leans on full-modal input as its differentiator, arguing most open Pro models are text-or-image only.
- vs. Xiaomi's own past: the V2.6 generation keeps the same pretraining architecture but reportedly doubles intelligence via far larger RL compute.
What the open-weight move means globally
Xiaomi is not a typical model lab — it is a hardware company that ships billions of devices a year. Open-sourcing MiMo-V2.6 means the weights are free for anyone to download, fine-tune, and embed, not just license through a paid API. For global developers, that lowers the barrier to building on a Chinese full-modal model without sending data to a Chinese server.
The cost framing is the real hook. If a frontier-level task genuinely costs around 13 cents through MiMo versus several dollars through a closed API, the economics of shipping AI features change for small teams and indie developers. The caveat is that "task cost" depends entirely on how you define a task; Xiaomi's number reflects its own measurement, and your mileage will vary with prompt style, context length, and retry rate.
There is also a strategic read. A device maker that owns both the silicon-to-showroom pipeline and the model can bundle intelligence into hardware at near-zero marginal cost — a position pure software labs cannot easily copy, and one that explains why a phone company is willing to give away frontier-grade weights.
Honest limitations
- The Artificial Analysis "46 points" and the cost ratios are Xiaomi's framing; we did not reproduce them, and benchmark methodologies differ across rounds.
- "Only open-source full-modal Pro model" is Xiaomi's claim; the field moves monthly, so verify before quoting it as permanent.
- The PFAS/MOF "Co-Scientist" example is an internal demo, not a published, peer-reviewed result.
- We did not benchmark latency, multimodal quality, or true per-task cost on our own workloads.
What readers can do now
- Download and test: pull MiMo-V2.6-Pro weights from Hugging Face and run the MiMo Desktop client or an OpenAI-compatible API endpoint.
- Stress the cost claim: run your own representative tasks and measure tokens spent versus a closed model to see if the 13-cent figure holds.
- Watch the ecosystem: Xiaomi's ¥60B / US$8.5B three-year AI commitment signals deeper on-device-model integration across phones, cars, and home devices.